Awesome

X-Mixup

Implementation of ICLR 2022 paper "Enhancing Cross-lingual Transfer by Manifold Mixup".

Structure

.
├── data                              # XTREME data and translation data
│   ├── xnli
│   │   ├── XNLI-MT-1.0
│   │   ├── XNLI-1.0          
│   │   ├── translate-test
│   │   ├── translate-dev
│   │   ├── translate-train-en  # back-translation data
│   ├── pawsx
│   │   ├── en (each language)
│   │   │   ├── train.tsv (only en)
│   │   │   ├── dev_2k.tsv
│   │   │   ├── test_2k.tsv
│   │   ├── translate-train        
│   │   ├── translate-test
│   │   ├── translate-dev
│   │   ├── translate-train-en
│   ├── udpos/panx
│   │   ├── en (each language) 
│   │   │   ├── train.tsv (only en)
│   │   │   ├── dev.tsv
│   │   │   ├── test.tsv
│   │   ├── translate-train-logits
│   │   ├── translate-test-token
│   │   ├── translate-dev-token
│   │   ├── translate-train-en-token
│   ├── squad
│   │   ├── train-v1.1.json
│   │   ├── translate-train
│   │   ├── translate-train-en
│   ├── mlqa
│   │   ├── dev
│   │   ├── test
│   │   ├── translate-test
│   ├── xquad
│   │   ├── test
│   │   ├── translate-test
│   ├── tydiqa
│   │   ├── dev
│   │   ├── translate-test
│   │   ├── translate-train
│   │   ├── translate-train-en         
├── scripts
├── xmixup
├── README.md
└── requirements.txt

Environment

pip install -r requirements.txt

Data

Prepare data before the training phrase:

• Step 1: Download XTREME data from XTREME repo (Note that we should keep the label of test set for evaluation).
• Step 2: Download other translation data from here.
• Step 3: Organize data following the Structure part.

Training & Evaluation

bash ./scripts/train.sh [pretrained_model] [task_name] [data_dir] [output_dir]

where the options are described as follows:

• [pretrained_model]: xlmr or mbert
• [task_name]: pawsx, xnli, udpos, panx, mlqa, xquad, tydiqa
• [data_dir]: data directory
• [output_dir]: output directory

Citation

Please consider citing our papers in your publications if this project helps your research

@inproceedings{yang2021enhancing,
  title={Enhancing Cross-lingual Transfer by Manifold Mixup},
  author={Yang, Huiyun and Chen, Huadong and Zhou, Hao and Li, Lei},
  booktitle={International Conference on Learning Representations},
  year={2021}
}